Location embedding

Use location embeddings to enrich prediction mappings.

We represent locations by embedding tiles into a vector. This technique is also known as loc2vec and was first described in this blogpost. An implementation in pytorch is described here, with a corresponding GitHub repository. We provide a tensorflow impelmentation.

Data generation

Tile generation

We generate tiles from the opentopomapp. Latitude and longitude coordinates for a certain region can be found using this tool. Tiles are generated by running python src/downloader.py.

The zoom level depends on you use case.

Training the network

The feature extractor consists of a pre-trained (imagenet) DenseNet121 followed by a 1x1 convolution layer, a dense layer and a final embedding layer (with linear activations).

The model can be trained using by running python src/main.py.

Visualizng the results

Results can be visualized by running jupyter notebook notebook/tensorboard_embedding_visualization.ipynb and running all cells. Afterwards the projections can be visualized in Tensorboard by running tensorboard --logdir logs/ in the terminal. Embeddings are found under the PROJECTOR tab.

Note: tensorboard projector documentation is written for tensorflow==1.*, whereas the model is trained in tensorflow==2.1. As such, one will have to create a separate virtual environment for the visualization (see requirements_tensorboard.txt)